5D Library › 回溯信念锁定、序贯痛点唤醒与马尔可夫升单自动机 · English

第八章 一个用户自己的核算器与一项预注册的盲法检验

给任何测算服务的四个问题

本章把全书变成四个问题,可以向任何测算服务——包括作者自己的——提出。它们不是营销或经营建议。

一、它的回溯断言能不能事先写下、事后核对? 泛泛的断言几乎不是证据;只有可以核对、事先记录的细节才是(第一、二章)。

二、它的升单价格与时机,是否远在用户的抗拒阈值之下、并在用户感受到结果之后? 推到阈值并不划算(第三章)。

三、它有没有冷静期与困境筛查? 没有护栏时,越强的唤醒使买家中处于困境者的比例越高(第四章)。

四、它的转化预测是否先分群、再建链,并说明了估计的不确定性? 一条链装不下两种人,稀有转移要上万位用户才估得准(第六章)。

用户自己的核算器

一位用户在一次测算之后,可以把测算师说过的回溯断言按种类数一数:泛泛的、半特异的、可以核对的,各命中几条、失误几条,再从自己的先验出发计算后验。

推导

# Chapter 8. [p15] A calculator a client can run after a reading.
# She counts the reader's retrodictive statements about her past by
# kind, as hits and misses, and starts from her own prior (here 0.01).
# The true likelihood ratios are those of p01 (generic h 0.95, b 0.80;
# semi-specific 0.90, 0.30; checkable detail 0.90, 0.02); the "felt"
# posterior uses the Forer base rate 0.20 for generic statements and
# ignores misses. Example: 6 generic hits and 1 miss, 2 semi-specific
# hits and 1 miss, no checkable detail.
import math

KINDS = {"generic": (0.95, 0.80, 0.20),
         "semi-specific": (0.90, 0.30, 0.30),
         "checkable": (0.90, 0.02, 0.02)}
COUNTS = {"generic": (6, 1), "semi-specific": (2, 1), "checkable": (0, 0)}
PRIOR = 0.01

def posterior(true):
    lo = math.log(PRIOR / (1 - PRIOR))
    for k, (hits, misses) in COUNTS.items():
        h, b, felt = KINDS[k]
        bb = b if true else felt
        lo += hits * math.log(h / bb)
        if true:
            lo += misses * math.log((1 - h) / (1 - b))
    return 1 / (1 + math.exp(-lo))

if __name__ == "__main__":
    for k, (hh, mm) in COUNTS.items():
        print("   {:14s} hits {}  misses {}".format(k, hh, mm))
    print("   posterior of real insight, true ratios : {:.4f}".format(
        posterior(True)))
    print("   posterior as it feels (Forer, no misses): {:.4f}".format(
        posterior(False)))
    print("\nReading: the same reading that feels conclusive leaves a")
    print("careful client almost where she started. Only checkable")
    print("details, written down before they are checked, can move her.")
#    generic        hits 6  misses 1
#    semi-specific  hits 2  misses 1
#    checkable      hits 0  misses 0
#    posterior of real insight, true ratios : 0.0090
#    posterior as it feels (Forer, no misses): 0.9990
#
# Reading: the same reading that feels conclusive leaves a
# careful client almost where she started. Only checkable
# details, written down before they are checked, can move her.

例子中,6 条泛泛断言命中、1 条失误,2 条半特异断言命中、1 条失误,没有可以核对的细节。按真实似然比,有洞察力的后验概率是 0.0090,比先验 0.01 还低;按感受(巴纳姆式基率、忽略失误),后验是 0.9990。同一次测算,感受上近乎确定,按证据计算则几乎原地不动。

一项预注册的盲法检验

假说 · 尚未验证

研究方案。 一套完整的测算引擎(断言生成规则、措辞模板)在招募前公开登记,此后不再改动。每位参与者提交出生资料;独立分析者用登记的引擎为该参与者及两位随机选取的其他参与者各生成一组回溯断言。参与者在不知道来源的情况下,挑出最符合自己过去的一组。主要结局:挑中为自己生成的那一组的比例;没有信息时为三分之一。检验:单侧精确二项检验,α=0.05\alpha = 0.05,把握度 0.90,备择假设为 0.45。由与作者无经济关系、不知道断言来源的独立分析者执行,结果不论正负都发表 (Carlson 1985)。

推导

# Chapter 8. [p16] A pre-registered blinded retrodiction test.
# Each participant gives birth data; an independent analyst runs the
# registered engine to write retrodictive statements for the
# participant and for two decoys (other participants' birth data). The
# participant, blind to the source, picks the set that fits her past.
# H0: hit rate 1/3. H1: 0.45. One-sided exact binomial test at 0.05,
# power 0.90: the smallest n and its critical number of hits.
from math import comb

def tails(n, p):
    """upper tails P(X >= k), k = 0..n, for X ~ Binomial(n, p)"""
    pmf = [comb(n, j) * p ** j * (1 - p) ** (n - j) for j in range(n + 1)]
    out, acc = [0.0] * (n + 2), 0.0
    for j in range(n, -1, -1):
        acc += pmf[j]
        out[j] = acc
    return out

def plan(p0, p1, alpha=0.05, power=0.90):
    for n in range(20, 2000):
        t0, t1 = tails(n, p0), tails(n, p1)
        k = next(k for k in range(n + 1) if t0[k] <= alpha)
        if t1[k] >= power:
            return n, k, t0[k], t1[k]

if __name__ == "__main__":
    n, k, a, b = plan(1 / 3, 0.45)
    print("   participants {}, reject H0 at {} hits or more".format(n, k))
    print("   size {:.4f}, power {:.4f}".format(a, b))
    n2, k2, a2, b2 = plan(1 / 3, 0.40)
    print("   for H1 = 0.40 instead: {} participants, {} hits".format(
        n2, k2))
    print("\nReading: a fair test of retrodiction is cheap: a few hundred")
    print("participants. What makes it fair is the decoys, the blinding")
    print("and the registration, not the size.")
#    participants 152, reject H0 at 61 hits or more
#    size 0.0468, power 0.9015
#    for H1 = 0.40 instead: 447 participants, 166 hits
#
# Reading: a fair test of retrodiction is cheap: a few hundred
# participants. What makes it fair is the decoys, the blinding
# and the registration, not the size.

以 0.45 为备择假设,需要 152 位参与者,命中 61 位或以上即拒绝原假设;实际的检验水准为 0.0468,把握度为 0.9015。若备择假设取更保守的 0.40,需要 447 位参与者。公平的回溯检验并不昂贵;使它公平的是诱饵、盲法与预先登记,而不是样本量。

小结

四个问题——断言能否核对、升单是否远离阈值、有没有护栏、转化预测是否分群——可以向任何测算服务提出。用户自己的核算器显示,一次「感觉很准」的测算,在证据上可能几乎没有移动先验。检验回溯能力的盲法实验只需 152 位参与者(备择 0.45)到 447 位(备择 0.40)。

参考文献

Carlson, Shawn. 1985. “A Double-Blind Test of Astrology.” Nature 318: 419–25.

Fan, Y. (2026). Retrodictive Belief Locking, Sequential Pain-Point Agitation and Markovian Upsell Automata: Bayesian Trust Avalanches, Sunk-Cost Commitment Ratchets and Cross-Domain Transition Matrices [回溯信念锁定、序贯痛点唤醒与马尔可夫升单自动机] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090186 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).